Inspur MetaBrain SD200 Ultra Packs 128 Domestic AI Chips for 2.8T Kimi K3 Under 5.85ms/Token
Inspur unveiled the MetaBrain SD200 Ultra supernode AI server at AICC2026, coupling 128 domestic AI chips with 8 TB of unified accelerator memory and 64 TB of host memory. The company claims it can run Moonshot AI's 2.8-trillion-parameter Kimi K3 on a single machine with token latency under 5.85 ms.
- 128 domestic AI chips, 8 TB accelerator memory and 64 TB host memory
- Kimi K3 at 2.8T parameters on one node, under 5.85 ms per token
- 0.69-microsecond memory latency, AllReduce about 3.5x faster
- Chip vendor undisclosed; claimed headroom for 10T-parameter models
Read next
Hardware